In-context learning can mitigate catastrophic forgetting; and what happens after we solve continual learning?

This talk will have two parts. In the first part, I will discuss our new results testing the hypothesis that in-context learning (ICL) can mitigate catastrophic forgetting -- because ICL enables more accurate posterior inference about the current "task" or "context", which can in turn enable more targeting weight updating. In the second part, I will discuss the implications of a world with widely deployed continual learning agents, and how it poses major challenges for AI evaluation and alignment -- many techniques assume a single static base model, and are not suited for dynamically changing models.